Introaied nancy2019 luengo

Document confidentiel –
ne peut être reproduit ni diffusé
sans l'accord préalable
de Sorbonne Université.
TEL and Artificial
Intelligence bring
to Education
1
Vanda.luengo@sorbonne-universite.fr

QUI JE SUIS ?
https://www.lip6.fr/
EDTECH masterhttp://sciences.sorbonne-
universite.fr
ANDROIDE http://androide.lip6.fr/

6
7 researchers
6 PhD students + 5 finished in 2019
2 postdoctoral researchers
1 contractual engineer
Several masters students
=> Currently we have two postdoctoral open
LIP6
MOCAH team

7
MOCAH: Current research
Cognitive diagnostic and learner
assessment
▪ In mathematics
▪ In physics
▪ In medicine
Adaptation, personalization and
feedback
▪ Context awareness
▪ Differentiated feedback to learners / teachers /
designers / institutions
▪ Highly interactive systems
▪ Massively multi user learning systems
▪ Ubiquitous environments
▪ Collaborative Learning games
Authoring tools
▪ Meta-design / Co-design
▪ For end-users (teachers)
Traces analysis
▪ Group detection, peer recommendation
▪ Knowledge pattern extraction (from forums,
interactions, etc.)
▪ Behavior pattern extraction
Merging symbolic and numerical
approaches
Applied and Multidisciplinary
research

10
But also
=> EDM 2021 will be in PARIS, Sorbonne Université

11
Today
This morning introduction to AIED domain
• Historical point of view to introduce the classical
models
Afternoon an example that I developped during 11 years
• Intelligent tutoring system in ortopedic surgery

13
Some questions
What is AIED?
What can it bring?
How?
How to connect AI to teaching and learning?
Risk: adapting to what is liked, not what is learned
Making an impact at the right level: teachers & students
level
Risk: more efficient administration, not more efficient learning
Ethics: what acceptable uses?
Risk: new ideas once the data is there

14
Goals today
Titre de la présentation
Not replacing the teachers: harness the strengths of AI to
empower them
Learning that is more:
Personalized
Flexible
Inclusive
Engaging
Respond to what is being learned, how it’s being learned, what
is felt…
The dilemma:
Decades of research: great ideas are there
(Cheap) technological devices are there
Practical applications… not so much
R. Luckin, W. Holmes, M. Griffiths & L. B. Forcier Intelligence
Unleashed – an Argument for AI in Education”, 2016

15
What is AIED?
30 years of research
AI + learning sciences (education, psychology,
neuroscience, linguistics, sociology) to promote adaptive
learning environments
Goal: opening the black box of learning
Methods:
Using theories
Building models (knowledge about the world)
Process with algorithm
Testing
Iterating

16
What models?
Domain model: what to teach, to learn
Learner model: students’ status (frequent update)
Achievements, skills, difficulties
Emotional state
Engagement (time on task)
Other models => metacognitive, emotional, social…
Pedagogical model: how to teach
Productive failures (mistakes are ok)
Appropriate feedback (which hints, when…?)
Testing knowledge (how?)

17
Classical AIED models => first period
(ITS)
Domain model
Pedagogical
model
Learner
model
Algorithms:
decision making
process
Learning content
individualized
xAdaptive learning environment

18Titre de la présentation
Domain model
Pedagogical
model
Learner
model
Algorithms:
decision making
process
Learning content
individualized
xAdaptive learning environment
Logs
Students’ actions,
Affect,
Speech…
Algorithms:
data analysis
Feedback
Updates
Updates
Open
Learner
Model
Classical AIED model => now
R. Luckin, W. Holmes, M. Griffiths & L. B. Forcier Intelligence
Unleashed – an Argument for AI in Education”, 2016

19
History of AIED = ITS
Precursors examples
SCHOLAR (Carbonell, 1970)
This program attempted to engage the student in a mixed initiative dialogue on
South American geography. The program and student communicated through a
sequence of natural language questions and answers. The tutor could both ask
and answer questions and keep track of the ongoing dialogue structure.
SOPHIE (a SOPHisticated Instructional Environment)
(Brown, Burton & deKleer, 1982)
It was a tutor for electronic troubleshooting that used its expert system to evaluate
the measurements students were making in troubleshooting a circuit.
GUIDON (Clancey 1982)
an ITS for teaching medical diagnostic problem solving first attempt to adapt a
pre-existing expert system into an intelligent tutor. It attempts to transfer expertise
to the students exclusively through case dialogues where a sick patient (the
'case') is described to the student in general terms.

20
History of ITS, before 1980’s
Scholar
SCHOLAR was a pioneering effort in the development of
computer tutors capable of handling unanticipated student
questions and of generating instructional material in
varying levels of detail, depending on the context of the
dialogue.
It was a mixed-initiative ITS: both the system and the
student could initiate conversation by asking questions.
Both the program's output and the student's inputs were
English sentences.

21
Tuteurs, Scholar
> approx what is the area of Brazil ?
S : 2 500 000 square miles
> wrong. Please indicate if the following statement is
correct: the area of Paraguay is approx 47 432 square miles
S : what is the area of Brazil
> the area of brazil is approx 3 600 000 square miles. Now
answer the question you didn't answer before.
S : false
> rephrase your statement. I can't understand it
S: incorrect
>very good. The area of Paraguay is approx 157 000
square miles. The language of Brazil is French. Right or wrong ?

22
Tuteurs, Sholar
> The eastern and central plain of Argentina is...
S : Chaco
> No.
The northern plain of Argentina is the Chaco. The eastern
and central plain of Argentina is the Pampas.
S : is the Chaco in Chile?
> No.
The Chaco is in Argentina, Paraguay and Bolivia, not in
Chile.
S : are the Andes a region of Chile
> Well, you are almost right. More specifically, a region of
Chile is the Southern Andes.

23
2AIED MODELS
=> DOMAIN MODEL

24
Domain model
Source: Wenger, Etienne (1987) Artificial Intelligence and Tutoring Systems: Computational and Cognitive Approaches to the
Communication of Knowledge
Domain to be conveyed to the student, i.e. the knowledge of the
experts or how experts performs the domain.
dynamic model of the domain knowledge and a set of rules by which the system can
"reason.“
Roots in expert systems research (such as medical diagnostic or electronic
troubleshooting systems) in some ITS it have the ability to generate multiple correct
sets of solutions, rather than a single idealized expert solution.
=> contains the concepts, rules, and problem-solving strategies of the domain to be
learned. It can fulfill several roles: as a source of expert knowledge, a standard for
evaluating the student’s performance or for detecting errors, etc.“

25
Scholar
the expert knowledge module : geography of South
America.
Represented in a semantic network whose nodes
instantiated geographical objects and concepts.
Statements like 'Tell me more about Brazil' just invoked a retrieval of facts stored in the
semantic network.
The real power of this representation schema comes by recognizing that it is possible to
answer questions for which answers are not stored.
it is not necessary to store in the semantic network that 'Lima is in South America‘.
the program must know about the attributes concerned, e.g. 'location' and 'capital', and in
particular, that if x is capital of v and y is located in z then x is in z: this is a rule of inference.
Carbonell, J.R, 1970, AI in CAI: An AI Approach to CAI, IEEE
Transactions on ManMachine Systems, V. 11, pp, 190

26
Domain model, another example
Cognitive tutors : lessons learned Anderson, J. R.; Corbett, A. T.; Koedinger, K. R.
& Pelletier, R. 1995

27
Domain model,
several approches
Several theories
A Comparative Analysis of Cognitive Tutoring and Constraint-Based Modeling, Mitrovic,
Koedinger and Martin, User Modeling 2003, pp 313-322
Constraint base
Olhson’ errors theory
Model tracing
Anderson ACT theory

28
Domain model, limits
KnowIedge elicitation and codification can be a very time-
consuming task, especially for a complex domain with an
enormous amount of knowledge and interrelationships of
that knowledge.
 investigating how to encode knowledge and how to
represent it in an Intelligent system remains the central
issue of creating an expert knowledge module.
=> the problem of ill-defined domains remains also a
research question

29
Domain model
Hybrid aproaches
Semantic (rules, semantic networks) + numerical
approaches
• combine the advantages of different approaches in order
to overcome their limitations.
• different approaches can be better suited for different
parts of the same task so as to offer common or
complementary tutoring services.

30
Domain model
Hybrid aproaches
in CanadarmTutor (a robotic arm deployed on the
international space station which has seven degrees of
freedom) there are
1. There are not a good solution (or hard to ellicitate)
2. no clear strategy for arm manipulation as it moves from
the original configuration to the targeted configuration.
=> Several kind of knowledge : procedural, spacial,
declarative…

31
Domain model
Hybrid aproaches
3 kinds of models
1. an expert system (Kabanza et al. 2005)
2. a cognitive model (Fournier-Vigier et al. 2008) => model tracing
3. the partial task model approach (Fournier- Vigier et al. 2009, 2012) => Mining
sequential rules
Canadarm Tutor provide assistance for different parts of
the arm manipulation task. The result is tutoring services
which greatly exceed what was possible to offer with each
individual approach, for this domain (Fournier-Vigier et al.
2009).
Fournier-Viger, P., Faghihi, U., Nkambou, R., Mephu Nguifo, E.. CMRules: Mining sequential rules
common to several sequences. In Journal Know.-Based Syst. Vol. 25(1), (2012) 63-76

32
Domain model
Hybrid aproaches
• Itemset
• Sequence
• Algorithm CMRule
2
MOOC EIAH
Fournier-Viger, P., Faghihi, U., Nkambou, R., Mephu Nguifo, E.. CMRules: Mining
sequential rules common to several sequences. In Journal Know.-Based Syst. Vol.
25(1), (2012) 63-76
{x y} => {z} means that at each occurrence of items x and y, we observe
an occurrence of item z (given support and confidence)
Figure from Fournier-Vigier 2012, p. 64

33
Domain model
Hybrid aproaches
“Canadarm Tutor provide assistance for different parts of
the arm manipulation task. The result is tutoring services
which greatly exceed what was possible to offer with each
individual approach”
Fournier-Viger, P., Faghihi, U., Nkambou, R., Mephu Nguifo, E.. CMRules: Mining sequential rules
common to several sequences. In Journal Know.-Based Syst. Vol. 25(1), (2012) 63-76

34
3AIED MODELS
=> STUDENT MODEL

35
Student model and diagnosis
Dynamic representation of the emerging knowledge and
skill of the students and it describes how to reason about
their knowledge
COMPUTER MODEL OF THE STUDENT ≠ STUDENT COMPUTER MODEL
Data about student How the student reason…
DIAGNOSIS = Process of inferring the current state of the
knowledge

36
Student Model
Source: Wenger, Etienne (1987) Artificial Intelligence and Tutoring Systems: Computational and Cognitive Approaches to the
Communication of Knowledge
According to Wenger, student models have three tasks :
• They must gather data from and about the learner. This data can be explicit -- asking
the student to solve specific problems -- or implicit -- tracking the students
navigation and other interactions and comparing them to information about similar
learner responses.
• They must use that data to create a representation of the student's knowledge and
learning process. The system then uses this model to predict what type of response
the student will make in subsequent situations, compares that prediction to the
students' actual response, and uses that information to refine the model of the
student.
• The student model must account for the data by performing some type of diagnosis,
both of the state of the student's knowledge and in terms of selecting optimal
pedagogical strategies for presenting subsequent domain information to the student.

37
Student Model
ALGEBRA TUTOR
Correct rules +
« bug » rules

38
Student Models
errors models examples with aproaches
enumerative reconstructive generative
enumerative ACTP, BUGGY
Extensible list of
errors
DEBUGGY, LMS :
combination of
enumerated errors that
allow the
reconstruction of
observed errors
MEMO-II :
List of enumerated
errors with a link to
enumerated
misconceptions
reconstructive PROUST:
conception
reconstructive from
intentions using a
library of plans error
ACM, PIXIE,
ADVISOR :
errors reconstructed
from a neutral
language with
primitive
Young & O’Shea
(1981) :
Incorrect procedures
that are reconstructed
from the explanation
of the nature of the
errors
generative Bonar & Soloway
(1985) :
abstract library errors
that have been
generated to explain
the origin of the
observed errors
REPAIR :
Errors generated by
repetition of
processing impasses
REPAIR/STEP, matz
(1982) :
Reduction of the
occurrence of errors
by “misslearning”
Source Wenger 1987, p. 348

39
FROM STUDENT BEHAVIOR
TO STUDENT KNOWLEDGE

40
Numérical approche example
Knowledge tracing
Knowledge tracing has become the dominant method of
modeling student knowledge.
Knwoledge tracing is a predictive model
See :
https://www.youtube.com/watch?v=y0jXUuChTsg
Corbett and Anderson 1995

41
Student model Knowledge tracing
approach

42
Knowledge tracing
in others words
At each successive opportunity to apply a skill, KT updates
its estimated probability that the student knows the skill,
based on the skill-specific learning and performance
parameters and the observed student performance
(evidence).
Reye showed that KT is special case of a DBN which assumes parameters do not
change across time slices.
=> the conditional independence graph of KT can be drawn as the following
Figure.

43
Student model, from data
ALGEBRA TUTOR

44
Student model
knwoledge represented
Source : Woolf 2009

45
STUDENT MODEL
CURRENT OBJECTIVES

46
Challenges for 2030
The student model
We envision that by 2030 user models for students will be
complex, not only representing what students know, do and have
abilities for, but other factors too.
For instance, user models will track when and how skills were learned and what pedagogies
worked best for each learner.
Moreover, user models will include information on the cultural
preferences of learners, their personal interests, learning goals,
and personal characteristics,
to select the optimal mix of learning environments, pedagogy, visualizations, and contexts that
maximize engagement, motivation and learning outcomes for each individual.
When the learner is part of a group, the model will make the
best compromise among the individuals who are part of the
group.
European network of excellence http://www.stellarnet.eu/

47
Challenges for 2030
the architecture
Most likely, by 2030 user model servers will be readily
available for education.
Servers are similar to generic user models in that they are
separate from the application and will not run as part of it.
User modelling servers will be part of local area networks
or wide area networks and serve more than one
application instance at a time

48
Challenges
1. sharing of learner models across learning systems
In the long-term, this trend may lead to a more integrated
and effective educational experience for students, across
their life-time of learning.
In the long-term, as the field gets better at developing,
refining, and exploiting
=> sophisticated multi-dimensional models of learners,
there is improved potential for tailoring each student’s
learning experiences to their educational needs.
M. C. Desmarais, R. S. J. d. Baker, A review of recent advances in learner and skill modeling in intelligent
learning environments. User Model User-Adap Inter (2012) 22:9–38

50
History of ITS 1980’s
Architecture : Pedagogical model
Source: Wenger, Etienne (1987), http://www.cse.msu.edu/rgroups/cse101/ITS/its.htm
Designs and regulates instructional interactions with the
student. Represent teaching strategies and includes
methods for encoding reasoning about the feedback
Teaching strategies : examples, analogies, ….
learning is viewed as successive transitions between knowledge states, the purpose of
teaching is accordingly to facilitate the student's traversal of the space of knowledge
states." (p. 365)

51
But…
More systems :
Inform the diagnosis to the human
Propose a direct feedback for each diagnosis.
If diagnosis=X them feedback=Y
=> Few systems propose an independent computer model
which models the decision of a pedagogical feedback
following a diagnosis.

52
Feedback in Cognitive Tutors
provide immediate feedback after each problem-solving
step.
The feedback uses rules and mal-rules of the student
model
Source : Anderson & all 1995 (Cognitive Tutors, Lessons Learned)

53
Pedagogical model : feedback
example
Rule that
gives the solutions

54
Feedback in Cognitive Tutors

55
Feedback in other kind of
model Tracing tutors (Andes)
Ref : VanLehn et all. 2005, The Andes physics Tutors, Lessons Learned

56
Feedback in other kind
of model Tracing tutors (Andes)
Three kinds of feedbacks
Flag feedback : red incorrect and green if it is correct
In order to give immediate feedback, entries relevant to solving the problem and the solution point for
the problem. In principle, this information could be provided by a human author instead of being the
solution graph file needs to contain the set of nonequation generated by Andes”.
Explications feedbacks (What's Wrong Help)
In order to implement What's Wrong Help, Andes needs only three sources of knowledge: the
knowledge base of error handlers, one solution point per problem, and one set of defined quantities
per problem
Hints to solve the problem (Next Step Help)
it gives a sequence of hints intended to accelerate learning.
First Implementation : The first version attempted to recognize the student’s plan for solving the
problem and hint the next step in that plan. The Andes1 Bayesian net was used for both plan
recognition and next step selection. But p is not consistent with expert human help.
the new version of Next Step Help does not try to recognize the student’s plan, but instead it engages
the student in a brief discussion that ensures the student is aware of the principle that is being used
to solve the problem, then selects a step the student has not yet done, and hints that step
=> Feedback
based in
domain
model.
=> Feedback
based student
model (the
solution graph
with errors)

57
Feedback models based in
Bayesian student models
Given a Bayesian student model, the next issue is how to use
the model to optimise the pedagogical actions of the intelligent
tutor.
Source : Mayo et al 2001. Optimising ITS Behaviour with Bayesian Networks and Decision Theory

58
three general approaches
Alternative Strategies
Diagnostic strategies
Decision-theoretic pedagogical

59
Alternative Strategies : optionally take the posterior
probabilities of the Bayesian network and use them as the
input to some heuristic decision rule.
ADELE (Ganeshan et. al., 2000). ADELE has a Bayesian network model of the domain
knowledge, but it uses a heuristic based on focus-of-attention to select the node in the
network about which to provide a hint.
Andes and SQL tutors was used also heuristics decisions rules at different levels

60
Diagnostic Strategies: The basic idea is to select actions
whose outcomes are likely to maximise the posterior
precision of some node in the network.
Millán et. all domain is test question selection, and questions are selected to maximise
the system’s certainty that the student has mastered the domain concepts. This strategy
has limited applicability outside of diagnostic tests.

61
Decision-Theoretic Pedagogical Strategies: select tutorial
actions that maximise expected utility.
While diagnosis is obviously an important component of expected utility maximisation,
it is only a secondary component. The primary consideration of an expected utility
calculation is the likely outcomes of the action, and their pedagogical utility
in CAPIT the expected utility of an action (e.g. problem selection) depends on the likely
outcomes of the action (e.g. how many errors are made).
The impact of the feedback could be calculated with some type of factors
In DTTutor, the action’s impact on many different factors related to the student (e.g. their
morale, etc) has an influence on expected utility.
Source : Mayo et all 2001. Optimising ITS Behaviour with Bayesian Networks and Decision Theory

62
BN Extension : influence diagram
or decision network RB
Use the value information to decide.
Inference of the utility based in a value
information
Three type of nodes
Chance nodes : represent random variable, like in clasical BN
Decision nodes : represent points where decisions maker has
a choice of actions
Utility nodes (or value nodes) : represent the utility function.
Function that maps from states to real numbers

63
Feedback, computer
considerationsCategory Differentiation
Modality Monomodale, multimodal
Adaptation Non adaptive, adaptive (macro adaptive, micro-adaptive)
Personalization Non-personalized, Personalized
Independency of the
model
Strong coupled, weakly couple, independent
Automatic Automatic, semi-automatic
Intelligence
Non intelligent
Intelligent
optimization
strategy
Linguistic
Alternative
Diagnostic
Decision Theory
Textual
Oral

64
Other numerical pedagogical
model
AI for Education - F. Bouchet

65
Perspectives pedagogical
model
« A single teaching strategy was implemented within each
tutor with the thought that this strategy was effective for
all students. However, students learn at different rates and
in different ways, and knowing which teaching strategy (…)
is useful for which student would be helpful. This section
suggest the need for multiple teaching strategies within a
single tutor so that an appropiate strategy might be
selected for a given student »
Woolf (2009, p. 133)
 Several statrategies
 Choose the better at the right time
=> DATA NOT ONLY FROM STUDENT BUT ALSO FROM
TEACHERS, TUTORS;;;

66
Perspectives
Assesment and feedback
[…] assessment instrument can assist the teacher in giving
good feedback to the student. However this takes time and
developing automatic marking and feedback systems for
formative assessment will assist with the scalability of
this proven effective pedagogical practice. Timely
feedback provides a motivating and important experience
for learners and the technology can reduce the over-
reliance of the teacher as the sole assessor.
Source stellar Document

67
CURRENT
AIED…
C

68
Current uses of AI for
Education
Personal tutors (ITS)
Intelligent support for collaborative learning (CSCL)
Intelligent virtual reality
68

69
Personal tutors :
intelligent tutoring systems
Individualized tutoring: an ideal but cannot scale… until now
Different pedagogical approaches:
Scaffolding learning: between support and challenge
Diagnosing procedural errors (BUGGY)
Helping learners to be in control: self-regulated learning
« learning how to learn » (MetaTutor) (Azevedo et al. 2013)
Different methods to get there:
Symbolical:
requires experts
uses models, ontologies…
Numerical:
requires data
uses self-training algorithm
(machine learning)
69
Hard to build
Hard to interpret

70
Supporting colloborative
learning
Collaboration helps learning:
Pair of students in online courses have higher learning outcomes than students learning
alone
Encourages reflexion
Caring about the group increases engagement
AI can help:
To form efficient groups with complementary skills
(Labarthe et al. 2016)
Identify efficient collaborative patterns (to help students or teachers)
Through pedagogical agents: tutor (AutoTutor), peer (learning by teaching (Betty’s Brain)(Biswas
et al. 2012))
Intelligent moderation: detect off topic discussions
70

71
Intelligent virtual reality
VR can help with learning:
Encouraging « what if » scenarios (simulation)
Visiting historical places
Enabling low-achieving students by shifting their self image
(Crystal Island)
AI can make virtual world « intelligent »:
DIRE QUELQUE CHOSE SUR LES GESTES ET PERCEPTIONS
Guiding learners to regulate their emotional status
Encouraging collaboration between learners
Applications:
Against bullying (FearNot!) (Vannini et al. 2011)
Peacekeeping scenarios (Traum et al. 2003)
71

72
FUTURE
AIED…
C

73
Future uses of AI for Education
?
Where is AI going?
What are the challenges?
Develop reliable indicators to track progress
More and more capture devices (biological data, eyetracking, speech recognition…)
Better understanding of the best teaching approaches and their context
More data collected + sharing
73

74
« A Renaissance in Assessment »
Just-in-time assessments
Today: LA can predict if a student will fail or drop-out
Tomorrow: how motivation and engagement varies
Tracking learning progress
Today: is the answer right/wrong?
Tomorrow: why? What type of mistakes? What emotional state?
Stealth assessments:
Today: short quizzes, final exam
Tomorrow: assessing while learning is happening (e.g. through a collaborative project)
74
(Hill and Barber, 2014)

75
New insights from learning
sciences
AIED will be more interdisciplinary than ever
Education neuroscience:
Today: Uncertain rewards can improve learning (Howard-Jones et al., 2014)
Tomorrow: calibrating rewards based on learner
Psychology:
Today: « growth mindset » in learners is more efficient than « fixed mindset » (static
intelligence) (Dweck, 2010)
Tomorrow: detecting student’s mindset and develop it
75

76
Lifelong learning partners
We learn more efficiently with another (Cole, 1996)
Today: learner-companion help stimulate student learning
in various ITS
Tomorrow: one assistant companion
Across app and devices
In and beyond school
Choosing the optimal resources when they are needed
76

77
Tackling unsolved issues
Achievement gap – a social issue:
Help those who need it the most with one-on-one tutoring
Making sure all benefits from AIED
Developing teacher expertise:
Reducing stress
Freeing time to do what humans do best with students (automatic grading, resource
recommendations…)
77

78
What about ethics?
Overall problem with data collection:
For what? For whom? Who decides?
Problem with algorithms:
What happens when AI goes wrong?
Who is responsible?
Sharing is necessary for AIED to be successful
Guaranteed anonymousness (privacy by design)
AIED encourages human behavior changes & to establish
relationships
Should every mistake be reported to the teacher?
Spy effect
78

79
Conclusion
AIED exists today, and will be more and more important in years
to come
It needs institutional support to spread:
Involve learners, teachers, parents in the co-design of the next AIED systems to meet their needs
Develop, evaluate… iterate
Spread data standards
Share data, but keeping the ethics in mind!
79AI for Education - F. Bouchet

80
Thanks to…
KIWI for the invitation to the summer school
François Bouchet for the first version of this
presentation
“Intelligence Unleashed – an Argument for
AI in Education” by R. Luckin, W. Holmes, M.
Griffiths & L. B. Forcier – on which this
presentation is loosely based on
You for your attention!
80

Introaied nancy2019 luengo

Empfohlen

Empfohlen

Weitere ähnliche Inhalte

Ähnlich wie Introaied nancy2019 luengo

Ähnlich wie Introaied nancy2019 luengo (20)

Mehr von Vanda Luengo

Mehr von Vanda Luengo (11)

Kürzlich hochgeladen

Kürzlich hochgeladen (20)

Introaied nancy2019 luengo